JSON Repair Tool

Markdown code fences and surrounding prose, truncation from a token limit, single quotes, unquoted keys, trailing and missing commas, // and /* */ comments, Python’s None, True and False, NaN and Infinity, curly quotes from a word processor, unescaped line breaks and tabs inside strings, hex and octal numbers, and several concatenated objects that were meant to be an array.

It tracks which brackets are still open as it reads, so when the document ends early it closes them in the right order. A half-written string or number at the very end is dropped. That makes it the usual fix for an LLM response that hit its token limit mid-object.

The repairs are listed with their line numbers precisely so you can check. The parser reads character by character rather than running find-and-replace, so a comma inside a string is never mistaken for a comma between two values — which is exactly how regex-based fixers corrupt data. If the input is already valid JSON, nothing is changed at all.

Because 007 as a number is 7, and that is almost never what the data meant. A leading zero signals an identifier — an SKU, a pin code, an extension — so it is quoted instead.

A value that was turned into a JSON string and then serialised again, so the whole document is one quoted string containing more JSON. It happens when data passes through a queue or a log field. It is technically valid JSON, so unwrapping it is the one change made to a document that already parses — and it can be switched off.

No. The repair runs in your browser, which matters here more than anywhere else on the site — broken JSON from a real system tends to contain real keys, tokens and customer records.